Pragya Jaiswal Email & Phone Number

Dublin, County Dublin, Ireland

County Dublin, Ireland

Dublin, County Dublin, Ireland

● Worked on a generalised model using Multi Class Object Detection with EfficientDet for monitoring cow activity (lying/standing/drinking etc) to inform on herd comfort & welfare. Achieved a total F1 score of up to 94.1% for our customers. Tools: Tensorflow.● Led project to support farms with obstructions (baffles). This project expanded our target market by 20%, and led to better model performance across existing customers, with a jump in F1 score from 50.24% to 85.24% for customers with baffles.● Built automated model maintenance pipelines, with integration tests, to support continued training of our models with minimal manual effort. We also enabled training & model evaluations with Sagemaker in a CI/CD lifecycle. Tools: Metaflow, AWS Wrangler, Parquet, Sagemaker, Mlflow, CircleCI.● Built a “Cow and Ear Tag Detection model” (CowNet) for identifying & tagging cows for near real-time inference on edge devices. This lightweight CNN & anchor based model uses MobileNet features, a Feature Pyramid Network with Depthwise Separable Convolution blocks and a box & class prediction head, and is trained on two classes that have significantly different scales. Worked with COCO & synthetic data to validate deep learning approaches & solve for challenges like occlusions, sizes, metrics (Fbeta) etc till real data is available.● Built a pipeline for individual Cow Identification by combining models that the team worked on. Used the “Cow and Ear Tag Detection model” (described above) with a tracking algorithm that uses a Kalman filter over detections. The results were fed to a Vision Transformer based Tag Reader to finally generate “cow walking” events. These tagged cow tracks were fed to a Lameness Pipeline that uses a Keypoint Detection model. The pipeline was designed with low coupling to enable retries from failures by saving intermediate states. Tools: Metaflow, AWS Step functions, Docker, OpenCV built from source, Parquet.

County Dublin, Ireland

● Analysis to understand how sunlight affects our customer insights every day due to our segmentation model struggling with saturated pixels.● Built and maintained our internal libraries to support daily timelapse videos for customers to ensure data quality of our insights, with full functional test coverage, runbooks and documentation to help on-call engineers.● Executed proof of concepts to find the technical feasibility and impact of informing customers if feed is out of reach from the cows. Used segmentation masks & thresholding. Tools: OpenCV, pandas, plotly & matplotlib.● Collaborated with internal and external stakeholders for manual data annotation, including cost estimates, contract discussions and training for the annotators.

Noida Area, India

● Applied traditional ML and shallow Neural Networks to a complex Content Moderation Problem that resulted in 55% automation by volume for our largest customer and 45% cost savings.● Built a framework for defining ML Pipelines to improve team’s efficiency and iteration speed by cutting down on boilerplate and providing best practices out-of-the-box like retrying from failure, processing data in chunks, and reusing artifacts for other models to cut on training time and deployment. The model saved the team days’ worth of work for every new model trained.● Collaborated with the engineering team for database-modelling and building the ML backend. This system interfaced with our engineering backend to serve near-real time predictions, allowing for human-in-the-loop interactions and automated review process. After 4 years, Squad still uses this system for delivering automation to its customers.● Built models using Spectrograms with CNNs on calls made through SquadIQ for KeyWord Spotting in bilingual audios. The goal was to help scale the business by automating quality assurance.● Solved for Noise Detection in audio recordings using attention pooling mechanism based CNNs trained using Mel Spectrograms with delta and delta delta features.● End to end Feature engineering from unstructured data present in the CRMs across the US Real Estate Industry, for lead prioritisation models.● Exploratory data analysis to find patterns and build a rule based algorithm to automate 100% of the work done by our manual team for filtering out leads that shouldn’t be contacted or seem to have low intent, thereby reducing human errors and increasing our conversions.

Pune Area, India

Responsible for automating the basic question generation process for the one way interviews on Monjin’s platform by employing natural language processing along with Machine Learning.

New Delhi Area, India

The work includes image processing and machine learning approach for the classification of chest X-rays as normal or abnormal. The algorithm has been tested on radiographs from China and has been shown to achieve an accuracy of 78 – 80% using Random Forest Classifiers.

New Delhi Area, India

Worked to develop a new tool to identify SILAC pairs in mass spectrometry data and quantify them.The results from the tool were integrated with those obtained from certain Database Search Engines and a scoring mechanism was developed to rank the final results obtained.

New Delhi Area, India

Epileptic Seizure PredictionWorked to develop a possible approach for detecting the onset of epileptic seizures well before its occurrence using Support Vector Machine on iEEG data.

New Delhi Area, India

Website development for Reva Industries Limited (MSME), Faridabad, India.Development of a web application - Thunderbird Address Book Synchronizer (TABS) and a Mozilla Thunderbird plugin - Thunder Sync, to synchronize and maintain a central address book as per the business requirements.

Bachelor Of Technology (B.Tech.), Information Technology And Mathematical Innovations

Activities and Societies: Autonomi - Robotics Society

Indian School Certificate Examination, English, Physics, Chemistry, Mathematics, Physical Education, 95.75%

Indian Certificate Of Secondary Education, English, Hindi, Environmental Education, Science, Mathematics, Computer Applications, Social Studies, 90.8%